Bird classification model (in progress)

CosmicBee · March 18, 2024, 2:18am

Hi there,

I’ve been working on a bird classification model. I’m hoping to get it onto my TinyML hardware but the RAM requirements are currently too large as I used a rather large input size. I’m working to further adjust it but I am fairly new to machine learning so it’s a process.

github.com

Timo614/machine-learning/blob/main/birds-tflite-test.ipynb

{
 "cells": [
  {
   "cell_type": "code",
   "execution_count": 9,
   "id": "dbe229d0-ebc1-48b6-a501-fb2063607363",
   "metadata": {},
   "outputs": [],
   "source": [
    "import tensorflow as tf\n",
    "import matplotlib.pyplot as plt\n",
    "from pathlib import Path\n",
    "import os\n",
    "import numpy as np\n",
    "import pandas as pd\n",
    "from tensorflow.keras.preprocessing.image import ImageDataGenerator"
   ]
  },
  {
   "cell_type": "code",

This file has been truncated. show original

I’m not sure if there are any downsides to my approach as my model is fairly small for the given number of outputs. I’m still adjusting it and I lost some accuracy with it early on before I refined my methods.

I included the heatmap and f1 scores from my model.

Apologies for the cross post from the discord show and tell just figured this is a larger community in retrospect (I can delete the discord message in that channel if needed from last week).

Thanks,

Tim

CosmicBee · March 26, 2024, 1:40pm

I refined my techniques (and came up with a few new ones) and ended up with an 85% accuracy model with int8 quantization at 96x96.

Fairly happy with the results as it appears it will fit on my microcontroller.

github.com

Timo614/machine-learning/blob/main/birds-tflite-test-96-96-84.ipynb

{
 "cells": [
  {
   "cell_type": "code",
   "execution_count": 2,
   "id": "dbe229d0-ebc1-48b6-a501-fb2063607363",
   "metadata": {},
   "outputs": [],
   "source": [
    "import tensorflow as tf\n",
    "from pathlib import Path\n",
    "import os\n",
    "import numpy as np\n",
    "import pandas as pd\n",
    "from tensorflow.keras.preprocessing.image import ImageDataGenerator"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,

This file has been truncated. show original

CosmicBee · May 28, 2024, 11:18pm

I restarted the work involved here to ensure I created better dataset splits for testing the end model. I’ve finished that process and documented it on my site: Bird Detection TinyML - Cranberry Grape | Cosmic Bee | Tim Lovett

The end model is 411 outputs 190,770 parameters 83.49% accuracy for test (kept pure) and 82% for test post int8 quantization to tflite.

github.com

Timo614/machine-learning/blob/main/birds/notebooks/birds_96x96_411_outputs_i87_full_relu6_post_decimation.ipynb

{
  "nbformat": 4,
  "nbformat_minor": 0,
  "metadata": {
    "colab": {
      "provenance": []
    },
    "kernelspec": {
      "name": "python3",
      "display_name": "Python 3"
    },
    "language_info": {
      "name": "python"
    }
  },
  "cells": [
    {
      "cell_type": "markdown",
      "source": [
        "Total params: 190,770\n",

This file has been truncated. show original

To date there has been little interest in this work but figured I’d document it in case anyone is interested in the future. I do fear I’m missing some significant issue with the model as to date I haven’t heard anything back from anyone I’ve told about it so apologies if this is all wasted effort. It does appear to work for me though when I upload images to confirm and I’d have thought the tflite accuracy would be terrible if it was a bad model.